To compare deep learning models of different architecture for automated lumbar spinal stenosis classification on MRI and benchmark their performance against radiologists and orthopedists. Lumbar spine MRI studies from Sep-2015 to Sep-2019 were retrospectively obtained. Exclusion criteria included previous spinal instrumentation, suboptimal image quality, post-gadolinium studies, and severe scoliosis. Axial T2-weighted and sagittal T1-weighted images were used. Studies were split into training/validation and test sets. An external test set of 100 studies was used. Training data were labelled by 4 radiologists using predefined gradings. Two models, CNN-based and transformer-based, were developed. Consensus labelling by two expert spine radiologists served as the reference standard. Test sets were labelled by 8 participants (2 general radiologists, 2 radiologists-in-training, 2 orthopedists, 2 orthopedists-in-training). Detection recall (%), interrater agreement (Gwet κ), sensitivity, and specificity were evaluated. 564 MRI lumbar spines were included (mean age = 52 ± 19[SD]; 302 women), with 464(82%) and 100(18%) for training/validation and internal testing, respectively. Both models showed high recall for all regions of interest (> 94%), similar to participants. Dichotomous classification (normal/mild vs. moderate/severe) by the CNN model, transformer model, and participants showed respective kappas for central canal 0.99/0.99/0.97-0.98, lateral recesses 0.98/0.94/0.81-0.94, and neural foramina 0.98/0.95/0.91-0.95 on internal testing (p < 0.001); for central canal 0.99/0.97/0.92-0.97, lateral recess 0.97/0.90/0.61-0.91, and neural foramina 0.99/0.94/0.87-0.93 on external testing (p < 0.001). The CNN model showed superior performance, and the transformer model showed similar to superior performance compared to clinicians for classifying lumbar spinal stenosis. These models could assist clinicians in report generation, surgical planning and education.

Artificial intelligence (AI) promises to accelerate and democratize medical imaging, yet low- and middle-income countries (LMICs) face distinct barriers to adoption. This perspective identifies those barriers and proposes an action-oriented roadmap. Insights were synthesized from a Johns Hopkins Science Diplomacy Hub workshop (18 experts in radiology, AI, and health policy) and a scoping review of peer-reviewed and grey literature. Workshop discussions were transcribed, thematically coded, and iteratively validated to reach consensus. Five interlocking barriers were prioritized: (1) infrastructure gaps-scarce imaging devices, unstable power, and limited bandwidth; (2) data deficiencies-small, non-representative, or ethically constrained datasets; (3) workforce shortages and brain drain; (4) uncertain ethical, regulatory, and medicolegal frameworks; and (5) financing and sustainability constraints. Case studies from Nigeria, Uganda, and Colombia showed that low-field MRI, cloud-based PACS, community-engaged data collection, and public-private partnerships can successfully mitigate several of these challenges. Targeted policy levers-including shared procurement of low-cost hardware, regional AI and data hubs, train-the-trainer workforce programs, and harmonized regulation-can enable LMIC health systems to deploy AI imaging responsibly, shorten diagnostic delays, and improve patient outcomes. Lessons are transferable to resource-constrained settings worldwide. Question How can LMICs overcome infrastructure, data, workforce, regulatory, and financing barriers to implement artificial-intelligence tools in clinical medical imaging? Findings Our multinational consensus identifies five obstacles and maps each to actionable levers: low-cost hardware, regional data hubs, train-the-trainer schemes, harmonized regulation, blended financing. Clinical relevance Implementing these targeted measures enables LMIC health systems to deploy AI imaging reliably, shorten diagnostic delays, and improve patient outcomes while reducing dependence on external expertise.

This study aims to address the scarcity of annotated Anterior Segment Optical Coherence Tomography (AS-OCT) datasets in ophthalmology by using Denoising Diffusion Generative Adversarial Networks (DD-GANs) to generate synthetic AS-OCT images to produce predictive models. The goal is to produce high-quality, diverse, and realistic data supporting the training of predictive models without data imbalance issues. AS-OCT images obtained from a tertiary referral hospital were used to train two DD-GAN models-one for healthy images and another for unhealthy AS-OCT images. The generated synthetic datasets were evaluated using Fréchet Inception Distance (FID) and Inception Scores to assess image quality. To further validate the synthetic data, two ResNet-50 models were trained separately on the real and synthetic datasets and evaluated on each other's test sets to measure performance comparability. Two synthetic datasets were created: a smaller set with 15.7 k images (6.8 k healthy, 8.9 k unhealthy) and a larger set with 100 k images (50 k each). The FID scores were 0.17 for healthy images and 0.23 for unhealthy ones, indicating high-quality synthesis. Inception Scores were 1.46 for healthy data and 1.55 for unhealthy data. ResNet-50 models trained on synthetic data achieved results comparable to models trained on real data. DD-GANs effectively generate realistic AS-OCT images, producing high-quality, balanced datasets that can address data scarcity and imbalance in ophthalmology. These synthetic datasets can enhance machine learning model development, advancing medical image analysis. Synthetic medical image generation provides significant advantages in protecting personal data privacy. By using artificially generated data instead of real data, patients' identities and confidentiality are safeguarded.

This study proposes DMCIE (diffusion model with concatenation of inputs and errors) to enhance binary brain tumor segmentation from multimodal MRI scans. Accurate voxel-wise tumor localization remains challenging due to variability in tumor size, shape, and imaging conditions, impacting clinical diagnosis and treatment planning. DMCIE employs a two-stage framework: a 3D U-Net first predicts an initial tumor mask from multimodal MRI inputs (T1, T1ce, T2, FLAIR), and an error map highlighting discrepancies with the ground truth is generated. This error map, concatenated with the original inputs, is refined through a diffusion model that iteratively corrects misclassified and boundary regions. The proposed DMCIE method was evaluated on the BraTS2020 dataset. Compared to the initial U-Net segmentation, DMCIE improved segmentation performance by +5.18% Dice and <math xmlns="http://www.w3.org/1998/Math/MathML"><mo>-</mo></math> 2.07 mm HD95 compared to the initial U-Net segmentation. It shows improvements in boundary accuracy and segmentation across diverse tumor shapes, and maintains spatial coherence, even in fragmented cases. DMCIE introduces an effective error-guided correction mechanism for binary brain tumor segmentation, using multimodal MRI data to enhance segmentation accuracy. By modeling and correcting segmentation errors during diffusion, DMCIE achieves anatomically precise and well-localized tumor segmentation.

This study aimed to develop and validate machine-learning (ML) models that integrate ultrasonic radiofrequency (RF) time-series signals with gray-scale image features for the preoperative differentiation of breast lesions classified as category 4A of the Breast Imaging Reporting and Data System. A dataset comprising RF signals, 2D ultrasound features, and pathological diagnoses from 130 BI-RADS 4A lesions (128 patients) was analyzed. Five ML models (logistic regression [LR], support vector machine [SVM], k-nearest neighbor [k-NN], and gradient boosting [GB]) were evaluated. Among 31 features (28 RF-derived and 5 2D image features), 6 key features were selected through feature selection. The LR model achieved the highest area under the curve (0.81, 95% confidence interval: 0.66-1.00), though no statistically significant differences were observed among models (DeLong test, p > .05). Artificial intelligence-assisted diagnosis improved accuracy across physician seniority levels (p < .05): junior (≤3 years: 52.28% versus baseline 27.28%), intermediate (4-10 years: 79.54% versus 45.46%), and senior (≥10 years: 81.91% versus 63.63%). The integration of RF time series and 2D features via LR demonstrates potential to reduce unnecessary biopsies by enhancing diagnostic precision, particularly for less experienced clinicians.

Photoacoustic computed tomography (PACT) reveals biological structures, pharmacokinetics, and physiological functions. Although a premium PACT system with many ultrasound (US) transducers delivers high-quality volumetric imaging, it suffers from high system costs and slow temporal resolution. Here, using a limited number of US elements, a hybrid diffusion model (HD-PACT) is demonstrated that enhances dynamic multiparametric (structural, functional, and contrast-enhanced) 3D PACT. Using just 256 out of the 1024 elements in a premier hemispherical US array for PACT, HD-PACT improves structural images acquired in different planes, organisms, and wavelengths. In functional imaging, HD-PACT enables 256-element PACT to observe hypoxia, pharmacokinetics, and angiogenesis during tumor progression. Lastly, HD-PACT is transferable to low-end PACT (only 128 US elements), where it dynamically captures contrast-free/enhanced organs, oxygen-perturbed brains, and cardiac dynamics with high spatiotemporal resolution in live animals. It is believed that HD-PACT will be valuable in oncology, cardiology, pharmacology, and endocrinology.

Study designRetrospective study.ObjectiveOsteoporotic vertebral compression fractures (OVCF) affect postmenopausal women, with 30-40% requiring surgical intervention after conservative treatment failure. This study developed a CT and MRI radiomics-based model to predict conservative treatment failure risk.MethodsWe retrospectively analyzed 154 postmenopausal women with OVCF (2016-2024), divided into successful (n = 86) and failed (n = 68) conservative treatment groups. Three-dimensional regions of interest were delineated, and quantitative features extracted using PyRadiomics. Feature selection employed Mann-Whitney U test, Spearman correlation, and LASSO regression. Clinical, radiomics, and combined models were constructed using eight machine learning algorithms with 5-fold cross-validation.ResultsAge and vertebral CT Hounsfield units were significant clinical predictors. From 3668 initial features, 16 key radiomics features were selected. LightGBM performed best for clinical models, while k-nearest neighbors excelled for radiomics models. In testing, the clinical model achieved AUC 0.684 (accuracy 0.71), radiomics model AUC 0.812 (accuracy 0.71), and combined model AUC 0.859 (accuracy 0.806). The combined model significantly outperformed individual models.ConclusionThe comprehensive CT and MRI radiomics-based model accurately predicts conservative treatment failure risk in postmenopausal women with OVCF. This tool enables early identification of high-risk patients and supports individualized treatment decisions, potentially guiding early surgical intervention for predicted high-risk cases.

The development of foundation models for brain MRI depends critically on the scale, diversity, and consistency of available data, yet systematic assessments of these factors remain scarce. In this study, we analyze 54 publicly accessible brain MRI datasets encompassing over 538,031 to provide a structured, multi-level overview tailored to foundation model development. At the dataset level, we characterize modality composition, disease coverage, and dataset scale, revealing strong imbalances between large healthy cohorts and smaller clinical populations. At the image level, we quantify voxel spacing, orientation, and intensity distributions across 15 representative datasets, demonstrating substantial heterogeneity that can influence representation learning. We then perform a quantitative evaluation of preprocessing variability, examining how intensity normalization, bias field correction, skull stripping, spatial registration, and interpolation alter voxel statistics and geometry. While these steps improve within-dataset consistency, residual differences persist between datasets. Finally, feature-space case study using a 3D DenseNet121 shows measurable residual covariate shift after standardized preprocessing, confirming that harmonization alone cannot eliminate inter-dataset bias. Together, these analyses provide a unified characterization of variability in public brain MRI resources and emphasize the need for preprocessing-aware and domain-adaptive strategies in the design of generalizable brain MRI foundation models.

This paper presents the FuzzyDistillViT-MobileNet model, a novel approach for lung cancer (LC) classification, leveraging dynamic fuzzy logic-driven knowledge distillation (KD) to address uncertainty and complexity in disease diagnosis. Unlike traditional models that rely on static KD with fixed weights, our method dynamically adjusts the distillation weight using fuzzy logic, enabling the student model to focus on high-confidence regions while reducing attention to ambiguous areas. This dynamic adjustment improves the model ability to handle varying uncertainty levels across different regions of LC images. We employ the Vision Transformer (ViT-B32) as the instructor model, which effectively transfers knowledge to the student model, MobileNet, enhancing the student generalization capabilities. The training process is further optimized using a dynamic wait adjustment mechanism that adapts the training procedure for improved convergence and performance. To enhance image quality, we introduce pixel-level image fusion improvement techniques such as Gamma correction and Histogram Equalization. The processed images (Pix1 and Pix2) are fused using a wavelet-based fusion method to improve image resolution and feature preservation. This fusion method uses the wavedec2 function to standardize images to a 224x224 resolution, decompose them into multi-scale frequency components, and recursively average coefficients at each level for better feature representation. To address computational efficiency, Genetic Algorithm (GA) is used to select the most suitable pre-trained student model from a pool of 12 candidates, balancing model performance with computational cost. The model is evaluated on two datasets, including LC25000 histopathological images (99.16% accuracy) and IQOTH/NCCD CT-scan images (99.54% accuracy), demonstrating robustness across different imaging domains.

Osteochondrosis encompasses a heterogenous group of pathologies affecting endochondral ossification in the growing child and adolescent. The cause of each disease is multifactorial, though many are often related to overuse injury, and may be epiphyseal, physeal, or apophyseal. Identification and treatment of this group of disorders is complex, thus this review aims to briefly describe common pathologies, their management, and highlight novel developments within the field. Machine learning as well as advanced diagnostic tools for more precise evaluation and prognostication of osteochondroses have been studied including perfusion MRI in Legg-Calvé-Perthes disease. Novel treatments include leukocyte-rich platelet-rich plasma (LR-PRP), which offer promising improvements in pain and function in Osgood-Schlatter disease. Surgical technique studies have begun to examine optimal operative management of Freiberg's disease. The osteochondroses are an often-self-limiting spectrum of pathologies affecting the physis in children and adolescents that may be managed with conservative treatment, though some require surgical intervention. Advances in imaging, prognostication tools, and treatment modalities support earlier and accurate diagnoses, as well as better informed treatment decisions.